Research On SSVEP Classification Algorithm Based On Deep Learning

Brain-computer interface technology builds a channel for direct information exchange between the brains and the external environments,makes it possible for the patients with brain diseases to communicate with the world.This thesis is dedicated to the brain-computer interface based on steady-state visual evoked potential(SSVEP),which has the advantages of less training time and high information transfer rate.It is an important research direction of the brain-computer interface.After years of research,improvements in stimulus paradigm design and signal processing algorithms have brought the system closer and closer to practical applications.However,most of the traditional signal processing methods are not desirable when they analyse SSVEP in a short time window.In recent years,deep learning has been effectively applied in the field of pattern recognition.In this thesis,the deep learning model is applied to SSVEP analysis in a short time window,and the research work is carried out from the depth mixed model and the semi-supervised ladder network respectively:1)Unsupervised learning networks can reconstruct signals by multiple greedy layers,which can effectively extract the intrinsic features of data.The restricted boltzmann machine(RBM)and autoencoder(AE)are two typical unsupervised learning algorithms.In this thesis,the both algorithms are improved for learning the features of multichannel SSVEP.Combined with supervised convolutional neural network(CNN),the deep mixed model is proposed for the SSVEP classification in short time window.Experiments on five subjects showed that the deep mixed model could achieve satisfying classification performance for SSVEP in short time window with an average accuracy of 90.56%.2)Supervised learning in the brain-computer interface requires collecting a large number of labeled samples,which takes time and labor.Considering the semi-supervised situation,a large number of labeled samples and a few labeled samples are used.In this thesis,the semi-supervised ladder network training method is used.The algorithm effectively combines the unsupervised reconstructed signal and supervised classification through the lateral skip connections.The experimental result shows that the semi-supervised ladder network is superior to other methods when there is a limited number of labeled samples.